Unveiling the Opacity: Exploring Rationality and Explainability in Tomato Disease Classification

Artificial Intelligence has advanced expeditiously over the past years and led to the production of highly complex and leading-edge models deployed to solve problems in a challenging environment. With high dependence on the machine made predictions, it has now become a necessity to be able to look transparently across the black-box and understand the features that influence it’s ability to make decisions. The paper contributes in understanding the rationality and explainability of the models in the classification of diseases in tomato fruit plants. We employed three models, Resnet101, EfficientNetB7, Vision Transformer (vit_base_patch16_224) to classify the disease present in tomato leaf. The opacity of these models are deciphered by implementing two Explainable Artificial Intelligence(XAI) methods, GuidedGradCAM and Integrated Gradients. The explainable algorithms help in unveiling and analyzing the factor that affects the prediction. According to the experimental results, EfficientNet-B7 demonstrates an exceptional level of accuracy, reaching a staggering 99.84% and classifies the disease based on the lesions. The Resnet101 also achieves a significant accuracy of 95.23%, whereas the vision transformer is able to achieve an accuracy of 77.76% only. The paper visualizes and draws a comparison between the two CNN-based models by evaluating their performance based on their validation loss and and accuracy. Interpreting the model’s intelligibility would help in strengthening the practitioner’s trust with the machine-driven decisions and ensure early detection and classification of the disease which would help in taking appropriate measures to prevent substantial crop and economical loss.